Simultaneous NODDI and GFA parameter map generation from subsampled q-space imaging using deep learning.

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Diffusion spectrum imaging (DSI) is crucial for generating neurite orientation dispersion and density imaging (NODDI) and generalized fractional anisotropy (GFA) parameter maps.
• Conventional DSI requires extensive q-space sampling, leading to clinically impractical scan times, especially for stroke imaging.

Purpose of the Study:

• To develop a deep learning-based method for simultaneous, accurate generation of NODDI and GFA maps from undersampled q-space data.
• To enable faster and more efficient neuroimaging in clinical settings, particularly for stroke patients.

Main Methods:

• A convolutional neural network (CNN) was trained to generate NODDI and GFA maps from 10x undersampled q-space data.
• The method utilized 48 DSI scans from stroke patients and healthy subjects for training, validation, and testing.
• Performance was compared against existing voxel-wise machine learning approaches.

Main Results:

• The CNN-based method demonstrated significant improvements in image quality metrics compared to previous machine learning techniques.
• Simultaneous generation of NODDI and GFA maps using a single network offered computational advantages.
• Predicted stroke functional outcomes using the generated maps showed minimal change (1-6% difference) compared to clinical evaluations.

Conclusions:

• Deep learning enables simultaneous estimation of NODDI and GFA parameters from highly undersampled q-space data.
• The proposed method offers a 10-fold reduction in scan time compared to conventional techniques.
• This advancement improves efficiency in neuroimaging for stroke assessment without compromising outcome prediction accuracy.